How to do inference with bounded constraints



I am trying to solve a specific problem and I am wondering IF pyro is the right tool for it and how to implement it.
I can simplify the problem to something like this:
I have 5 continuous variables: A, B, C, D, E, but A = 2*B+C and C = D+E
So, they are only 3 independent variables (B,D,E).

I know that B,D,E are gaussian distributed (so give numbers, let say that the prior is µ=0.5 and sigma=0.5), and all three must be >0.
and I know that A=1 (it’s a constraint) and that B is in [0,0.5]
how can I compute the posterior on B,C,D,E for this problem that will fit my constraints

thanks a lot for your help, as I don’t know how to start coding this.


Hi @loicus, currently truncated normal distribution (which has positive support) is not yet implemented in PyTorch and Pyro. I think that you would need to hack around a bit by using other priors instead. To find the posterior, you can use SVI or MCMC as in bayesian regression tutorial.


Are there plans to implement truncation (both max and min) for the different distributions? It seems like an important feature for practical appliations.